Anti-Symmetric Molecular Graph Learning Approach With Residual Adaptive Network Based Fuzzy Inference System for Lethal Dose Forecasting Problem.

In recent times, graph neural networks (GNNs) have become essential tools in molecular graph learning, due to its ability to model intricate structural dependencies. Despite their success, recent research has shown that GNNs still face significant limitations, in capturing long-range dependencies and global structural information. One of the central issues is the over-squashing problem, where information from distant nodes is excessively compressed into fixed-size node representations. This leads to poor information propagation; as a result, ultimately degrading the model's performance-particularly in complex tasks such as lethal dose forecasting, where both local chemical substructures and global molecular topology play vital roles. To overcome these limitations, we propose a novel anti-symmetric fuzzy-enhanced graph learning (ASFGL) model. Generally, our model integrates two key components: an anti-symmetric transformation module and a residual adaptive neuro-fuzzy inference system (ANFIS). The anti-symmetric transformation is designed based on stable graph ordinary differential equations (ODE); thus, ensuring a non-dissipative and stable propagation of information across multiple graph layers. This mechanism effectively mitigates the over-squashing issue, therefore, allows our model to better capture long-range dependencies in a stable manner. Complementarily, the ANFIS module employs bell-shaped membership functions to support robust and interpretable learning; as a result, enabling adaptive rule-based reasoning that refines the molecular representations learned from the graph structure. By combining these modules, the ASFGL model bridges local message passing and global structural awareness, yielding expressive molecular embeddings well-designed for toxicity prediction problems. We evaluate our proposed ASFGL model on different benchmark molecular datasets, where it consistently outperforms state-of-the-art GNN-based architectures in terms of MAE/RMSE evaluation metrics, particularly in scenarios requiring deep representation learning over large interactions. These results highlight the efficacy of integrating anti-symmetric dynamics and fuzzy inference systems in advancing molecular property prediction and overcoming foundational challenges in GNN design.